Cognitive Design Computing John S. Gero1 1Department of Computer Science and School of Architecture, University of North Carolina at Charlotte and Krasnow Institute for Advanced Study, George Mason University [email protected] This talk describes the foundational concepts of cognitive design computing and then presents some examples. Cognitive computing is concerned with modeling human cognition computationally and using that model as the foundation for constructing computer models of design activities. Human cognition is based on perception, learning and adaptation. Here we present human cognition in terms of situated cognition - cognition involving interaction with an environment. The talk briefly introduces a set of principles for cognitive design computing founded on the three concepts of interaction, constructive memory and situatedness. It then presents two examples of applications of this approach. INTRODUCTION Computational design tools aim at encoding knowl- edge and making it available in an objective man- ner to the designer. There is an assumption behind this approach, namely that all the knowledge is ob- jective, ie, independent of the user. Examples of objective knowledge include stress analysis, deter- mining the position of the sun, thermal analysis of a building and methods such as linear programming in optimisation. Such objective knowledge tends to be deductive in nature. The most powerful exam- ples of deductive knowledge are those based on ax- ioms from which subsequent theorems have been developed. These theorems map onto the behaviour of the world. In addition, there is a category of knowledge that is based on induction (ie, knowledge learned from examples without causality). Even in- ductive knowledge is treated as objective, ie, inde- pendent of the user. For example, a layout algorithm that utilises some heuristics is used as if there were tion of computational tools. It has provided the ba- causality encoded. This has served design comput- sis of transferable skill development in users. As our ing well. It has allowed for the widespread distribu- knowledge of the world has improved, so we have Keynotes - Volume 1 - eCAADe 35 | 37 been able to update and adapt the knowledge in represented in a CAD software system. Figure 1(a) these programs. shows the screen image of a floor layout. Simply This talk presents an approach that aims to extend looking at the drawing of the floor layout gives no in- our understanding of what kinds of knowledge we dication of how it has been encoded. The darkened can expect our computational tools to have and how line is the single polyline representation of the out- systems that have a range of kinds of knowledge line obtained by pointing to a spot on the boundary, might perform differently. but that representation could not be discerned from the image. Figure 1(b) shows exactly the same out- FIRST-PERSON VERSUS THIRD-PERSON line but it is encoded differently, as indicated by the KNOWLEDGE darkened polyline obtained by pointing to the same spot. We call such objective knowledge third-personknowl- The issue here is one of interpretation that is of- edge in that the person who produced the knowl- ten missing in computation. A common assumption edge is not required to be there when that knowl- is that the external (and even the internal) world is edge is used by another person. For example, even there to be represented, ie, that in some sense it has though Newton is dead his laws continue to work only one representation. This misses an important fine. Whilst it is clear that much of human knowl- step: namely that of interpretation, which depends edge is of this third-person kind, in the sense de- on the viewer not on the underlying external rep- scribed above, there is a category of everyday knowl- resentation. Before anything can be represented it edge that depends on the person rather than deduc- needs first to be interpreted and it is this interpreta- tion. This kind of knowledge develops through the tion that is represented. This is an example of first- interaction of the individual and their world and as person interaction with the external world that re- a consequence is personal knowledge. It is called sults in first-person knowledge about the world. first-person knowledge. This class of knowledge is How can we build computation systems that en- sometimes inappropriately encoded as third person code first-person as well as third-person knowledge? knowledge and when done so often causes the mis- To do this we rely on concepts from cognitive science match between the experience of the person who and in particular a branch called situated cognition. coded the knowledge and a subsequent user of that Using those concepts we can produce a branch of knowledge. computing called cognitive computing that is a closer A simple example of such encoding of personal analog to how the mind works than general comput- knowledge can be seen even in the way objects are ing. Figure 1 The same image has different encodings (a) and (b) that depend on the individuals who created them rather than on any objective knowledge. 38 | eCAADe 35 - Keynotes - Volume 1 SITUATED COGNITION signing, the computational systems are unchanged Cognitive science is concerned with understanding by their use. This makes sense if the system only em- and representing structures and processes in the bodies third-person knowledge. However, the use of mind (as opposed to brain science which is con- any system, whatever kind of knowledge it contains, cerned with understanding and modeling structures generates first-person knowledge in the system that and processes in the brain). Situated cognition is uses it. This is the first fundamental distinction be- concerned cognition that is embodied and a conse- tween traditional design computing and situated de- quence of embodiment is an increased focus on the sign computing. For example, an optimization ap- interaction between the computational system and proach used in the design of layouts could produce its environment (Clancey, 1997; 1999; Gero, 1998; first-person knowledge in the form successful strate- Suchman, 1987). gies. These could then be used next time the opti- Situated cognition is founded on three ideas: mization program was used for layouts. 1. Interaction: knowledge comes from encod- Principle of Ordered Temporality: When you do what ing and through interaction; in particular you do matters first-person interactions with representations, In traditional design computing when you carry out which includes the expectations of the per- a computation plays no role in the result that is pro- son or system carrying out the interpretation duced. Again, this makes sense if the system only (Agre, 1997; Smith and Gero, 1998; Zhang, embodies third-person knowledge. However, if the 1997); system embodies first-person and generates first- 2. Constructive memory: constructive memory, person-knowledge as it runs, the chronology of the which is concerned with memory as a pro- system’s use affects what is used and what is learned. cess to generate a memory cued on a demand As the boundary condition for this principle consider to have such a memory rather than a recall that if A is carried out before B, then A cannot make of elements in a location (Bartlett, 1932/1977; use of any knowledge acquired through the execu- Dewey, 1896; Gero, 1999; Rosenfeld, 1988; tion of B, but B can make of knowledge acquired von Glaserfeld, 1995); through the execution of A. 3. Situatedness: situatedness is concerned with being in a particular place at a particular time Lemma of Experience: What you did before affects and how the world is viewed by each indi- what you do now vidual from that place at that time (Suchman, As a consequence of principles 1 and 2, we can state 1987). this lemma: What a system did before affects it does now. This is one definition of experience. Experience has the potential to guide future actions. The effect PRINCIPLES FOR COGNITIVE DESIGN COM- of this is that cognitive design computing systems are PUTING not static systems but are dynamic in terms of their behavior. This concept can be applied recursively, so We can develop a set of principles that form the foun- that previous experiences are used to produce cur- dation for the development of cognitive design com- rent experiences. puting. Principle of Locality: Where you are when you do what Principle of Effect: What you do matters you do matters The implication of this principle is that actions pro- First-person knowledge includes not only what hap- duce effects in terms of the production of first-person pened and when it happened but also where it hap- knowledge. In traditional uses of computers in de- Keynotes - Volume 1 - eCAADe 35 | 39 pened. The system has to be at the ”right” place at Gero, JS 1998, ’Conceptual designing as a sequence of the ’right” time for unique events to be interpreted. situated acts’, in Smith, I (eds) 1998, Artificial Intelli- gence in Structural Engineering, Springer, Berlin, pp. Principle of Interaction: Who and what you interact 165-177 with matters von Glasersfeld, E 1995, Radical Constructivism: A Way of Knowing and Learning, The Falmer Press Interaction is one of the distinguishing characteris- Rosenfield, I 1988, The Invention of Memory, Basic Books, tics of a cognitive design computing system. With- New York out interaction there is no potential to produce first- Smith, G and Gero, JS 2002, ’Interaction and experience: person knowledge. There are two sources of inter- Situated agents and sketching’, in Gero, JS and Bra- action for a system: other computer programs and zier, F (eds) 2002, Agents in Design 2002, Key Centre users of the system. Conceptually there is no differ- of Design Computing and Cognition, University of Sydney, Australia, pp.